11-19 更新 实时
首页/ 深度/ 正文

初创公司借AI推进数学全面形式化

深度 来源:IEEE Spectrum AI 2026-11-18 21:00 0 阅读 机器翻译
初创公司借AI推进数学全面形式化

庞大的定理:AI 蜂群验证 15,000 页 - IEEE Spectrum

初创公司借助 AI 在形式化全部数学的道路上取得进展

致技术内幕人士

致技术内幕人士

IEEE Spectrum 是 IEEE 的旗舰出版物——IEEE 是全球最大的致力于工程与应用科学的专业组织。我们的文章、视频和信息图向读者介绍技术、工程与科学领域的最新进展。

创建账户,享受更多免费内容和权益

保存文章以便稍后阅读需要 IEEE Spectrum 账户

The Institute 的内容仅对会员开放

下载完整 PDF 期刊仅限 IEEE 会员

下载本电子书仅限 IEEE 会员

访问 Spectrum 数字版仅限 IEEE 会员

关注话题是 IEEE 会员专享功能

加入全球最大的致力于工程与应用科学的专业组织,即可访问 Spectrum 的全部文章、存档、PDF 下载及其他权益。进一步了解 IEEE →

加入全球最大的致力于工程与应用科学的专业组织,即可获得本电子书以及 IEEE Spectrum 的全部文章、存档、PDF 下载及其他权益。进一步了解 IEEE →

访问数千篇文章——完全免费

创建账户即可获得专属内容与功能:保存文章、下载合集、发表评论——全部免费!如需完整访问权限与更多权益,请订阅 Spectrum。

初创公司借助 AI 在形式化全部数学的道路上取得进展

Axiom Math 公布两项成果,展现 AI 辅助验证的广度与深度

Benjamin Skuse 是一名自由撰稿人,撰写各类科技题材。他拥有应用数学博士学位。

10 月 25 日周日,当 Axiom Math 的 AI 集群正为有限单群分类(CFSG)——俗称“庞大定理”——那份 15,000 页证明的自动验证做最后收尾时,该公司的创始数学家 Ken Ono 除了旁观并回味这一成就之外,无事可做。

作为数学家的儿子,他从小身边就围绕着 CFSG 证明的几位缔造者——这一证明被普遍认为是人类写下的最长数学证明。如今,他正在为形式化验证他父亲和同行们所创造的成果贡献力量。“我看着团队倒计时直至完成验证,我的妻子就陪在我身边,”他回忆道。“天啊,我无法形容自己当时有多激动。”

圣安德鲁斯大学群论学家 Colva Roney-Dougal(未参与此项工作)在谈到 CFSG 的形式化时表示:“我一直把它当作一个极难形式化的典型例子,因为我原以为它所需要的那种计算量根本无法实现,”她说。“所以当我听到这一消息时,我真的非常非常惊讶。”

这一成果,连同同期完成的 7500 页 Stacks 项目的形式化验证,代表了一项巨大的 AI 工程成就,并为全部数学的数字化奠定了一些基础。

为何如此轰动?CFSG 关注的是有限单群。所谓群,是指带有满足特定规则的运算(如加法或乘法)的元素集合;「单」意味着无法进一步约化或简化,「有限」则区别于包含无限多个元素的群。例如,一副牌的所有排列就构成一个有限单群。

近 200 年前,Évariste Galois 奠定了群论的基础。此后,众多数学家将有限单群划分为 18 个族(外加 26 个不属于任何族群的散在单群)。CFSG 的证明确认了这一清单的完备性。

但这一看似简单的任务,只有汇集远超任何单个数学家所能完全吸收的成果与论证才能完成——这些内容来自约 100 位作者在半个世纪里撰写的 500 多篇期刊论文。因此,将这张相互关联的知识之网转化为代码并验证其正确性,带来了诸多重大的工程挑战,甚至超出了 Axiom Math 今年早些时候在证明形式化工作中所克服的难题。

据主导 CFSG 形式化工作的 Axiom Math 的 Niels Voss 称,其中最困难的一点就是证明本身的长度。CFSG 的证明超过 15,000 页密集的数学内容,包含众多定理、引理和定义,而只有当所有内容都实现机器可读之后,才能对其中的任何一项进行检验。

要绕开这一点,需要一些巧思。“我们处理了来自 CFSG 的全部文献,并在能够着手之前先梳理出那个非形式化证明的网络结构,”Voss 说。“随后我们把 CFSG 的陈述拆解成大约 100 个里程碑——沿途可被独立验证的标志性结果——用它们来指导形式化工作并保持方向不偏。”

Roney-Dougal 说,她对这一结果印象非常深刻,尤其欣赏这个形式化证明找出并澄清了原始人类证明中的若干错误。她认为,这一结果对于形式化群论领域的任何现代研究都至关重要,比如她自己理解有限单群行为的工作。

但 Roney-Dougal 说,由于验证是自动完成的,如今很少有数学家能够从这一证明中提取信息并与学界分享,而新一代数学家或许“错过了一个通过亲手做这件事来获取知识的机会”。

仍需要成堆的巧思

就在 CFSG 形式化成果公布仅九天后,Axion Math 披露它也已自动形式化了 Stacks 项目。Stacks 项目由哥伦比亚大学数学家 Aise Johan de Jong 于 2005 年发起,是一部规模庞大、开源的协作式数学教科书与参考著作,聚焦于高度抽象的代数几何与代数栈(对现代弦论和数论至关重要的几何空间)。

CFSG 是一个单一(但极其庞大)的标志性定理,而 Stacks 项目本质上是一个结构良好的文库,由数千个相互关联的定义、构造和引理组成,把整个现代代数几何汇聚在一起。

最终完成的 Stacks 形式化仅有 1600 万行代码(相比之下 CFSG 有 4300 万行代码,大致与 Linux 内核相当),但它给 Axiom Math 的工程师带来了另一种挑战:需要在数千个相互关联的定义和定理之间实现忠实性与一致性,该项目负责人、Axiom Math 的 Vasily Ilin 说。

这要求团队设计一套分层的 AI 智能体系统,让承担专门职责的智能体并行工作,使整个库中的每一个定义保持一致,同时通过多智能体审计、与来源交叉核对以及严格测试来检查忠实性。这确保了每个要素和任务在整个项目中保持一致。

Stacks 形式化是一个高效、通用、可复用且可维护的形式化基础概念库,支撑着现代数学的许多内容,其影响可能十分深远。如今,任何人或任何机器人都可以访问形式化的 Stacks 库,并在其上构建新的形式化数学。而 CFSG 与 Stacks 两项形式化成果合在一起,代表着朝着形式化全部数学迈出的重要一步。“它们展示了形式化必须扩展的两种方式:一个是非常长的证明,另一个是非常庞大的相互关联的理论体系,”Ono 说。

Ono 和 Voss 认为,形式数学成为核心的未来拥有诸多优势:为定理和证明盖章认证、在专业知识随时间流失时保存知识,以及建立一个共享库作为创造新数学的基础。当全部数学都被形式化时,他们对人类所扮演的角色也持乐观态度,将形式化视为人类以新颖且富有创意的方式使用和理解数学的跳板。“想到数学的未来可以是什么样,我们难道不感到兴奋吗?”Ono 问道。“我为此感到兴奋。我真的非常兴奋。”

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Startup Makes Headway Towards Formalizing All of Math With AI)

Enormous Theorem AI Swarm Validates 15,000 Pages - IEEE Spectrum

Startup Makes Headway Towards Formalizing All of Math With AI

FOR THE TECHNOLOGY INSIDER

FOR THE TECHNOLOGY INSIDER

IEEE Spectrum is the flagship publication of the IEEE — the world’s largest professional organization devoted to engineering and applied sciences. Our articles, videos, and infographics inform our readers about developments in technology, engineering, and science.

Enjoy more free content and benefits by creating an account

Saving articles to read later requires an IEEE Spectrum account

The Institute content is only available for members

Downloading full PDF issues is exclusive for IEEE Members

Downloading this e-book is exclusive for IEEE Members

Access to Spectrum 's Digital Edition is exclusive for IEEE Members

Following topics is a feature exclusive for IEEE Members

Join the world’s largest professional organization devoted to engineering and applied sciences and get access to all of Spectrum’s articles, archives, PDF downloads, and other benefits. Learn more about IEEE →

Join the world’s largest professional organization devoted to engineering and applied sciences and get access to this e-book plus all of IEEE Spectrum’s articles, archives, PDF downloads, and other benefits. Learn more about IEEE →

Access Thousands of Articles — Completely Free

Create an account and get exclusive content and features: Save articles, download collections, and post comments — all free! For full access and benefits, subscribe to Spectrum .

Startup Makes Headway Towards Formalizing All of Math With AI

Axiom Math announces two results demonstrating the breadth and depth of AI-assisted verification

Benjamin Skuse is a freelance writer of all things science and technology. He holds a Ph.D. in applied mathematics.

A dot with thousands of plotted points radiating from its center.

As Axiom Math’s AI swarm was putting the final touches on the automatic verification of 15,000-page proof of the Classification of Finite Simple Groups (CFSG), colloquially referred to as the ‘enormous theorem’, on Sunday 25 October, the company’s Founding Mathematician Ken Ono had nothing to do but watch and reflect on the achievement.

As the son of a mathematician, he had grown up surrounded by some of the architects of the CFSG’s proof, widely considered to be the longest mathematical proof ever written by humans. Now, he was contributing to formally verifying what his father and peers had created. “My wife was with me as I watched the team countdown to completing the verification,” he recalls. “God, I cannot tell you how emotional I was.”

University of St Andrews group theorist Colva Roney-Dougal (who was not involved in the work) says of formalizing CFSG: “I held it up as an example of something that was going to be incredibly hard to formalize because I had thought that the amount of computing it was going to require was just unattainable,” she says. “So when I heard the announcement, I was really, really surprised.”

This result, along with the concurrent formal verification of the 7500-page Stacks project, represent a huge AI engineering achievement and lay some of the groundwork for digitizing all mathematics.

Why the hoo-ha? The CFSG concerns finite simple groups. These are groups (sets of elements combined with an operation, like addition or multiplication, that follows specific rules) that are simple (in other words, cannot be reduced or simplified further) and finite (as opposed to groups containing an infinite number of elements). For example, all the permutations of a deck of cards form a simple finite group.

Évariste Galois laid the foundations for group theory almost 200 years ago. Finite simple groups have since been classified by myriad mathematicians into 18 families (plus 26 sporadic simple groups that do not belong to any of these families). The CFSG’s proof confirms that this list is complete.

But this simple task was only achieved by bringing together results and arguments that far exceed what any individual mathematician can fully absorb from more than 500 journal articles written by around 100 authors over a half-century period. Translating this interconnected web of knowledge into code and verifying its correctness therefore posed a number of significant engineering challenges, even beyond those overcome in earlier proof formalizations by Axiom Math this year.

One of the hardest, according to Axiom Math’s Niels Voss, who led the CFSG formalization, was the sheer length of the proof. Over 15,000 pages of dense mathematics, the CFSG proof contains multiple theorems, lemmas, and definitions none of which can be checked until everything has been made machine-readable.

To get around this required some ingenuity. “We processed all the literature from the CFSG, and traced the network structure of the informal proof before we were able to start working on it,” says Voss. “We then broke up the statement of the CFSG into about 100 milestones—landmark results along the way that can be independently verified—which we used to guide the formalization and stay on track.”

Roney-Dougal says she is very impressed with the result, and particularly appreciates how the formal proof finds and clears up several mistakes in the original human proof. She thinks this result is critical to formalizing any modern research in group theory, such as her own work understanding how simple finite groups behave.

But, Roney-Dougal says, because the verification was done automatically, there are few mathematicians capable of extracting information from the proof to share with the community, and that a new generation of mathematicians may have “missed a chance to gain knowledge from the doing of it”.

Stacks of ingenuity required

Just nine days after the CFSG formalization announcement, Axion Math revealed it had automatically formalized the Stacks project, too. Started in 2005 by Columbia University mathematician Aise Johan de Jong, Stacks is a massive, open-source collaborative mathematics textbook and reference work focused on highly abstract algebraic geometry and algebraic stacks (geometric spaces that are vital to modern string theory and number theory).

Whereas CFSG is a single (huge) landmark theorem, the Stacks project is essentially a well-structured library of thousands of interlocking definitions, constructions, and lemmas bringing all of modern algebraic geometry together.

The final Stacks formalization stands at a mere 16 million lines of code (compared to CFSG’s 43 million lines of code, roughly on a par with the Linux kernel), but it presented a different challenge to Axiom Math engineers: needing to achieve faithfulness and coherence across thousands of interlocking definitions and theorems, says Axiom Math’s Vasily Ilin, who led the project.

This required the team to design a hierarchical system of AI agents with specialized responsibilities working in parallel, keeping every definition aligned across the library, while also checking faithfulness with multi-agent audits, cross-checks against sources, and hard tests. This ensured that every element and task aligned across the entire project.

A highly efficient, general, reusable, and maintainable formal library of foundational concepts underpinning much of modern mathematics, the impact of the Stacks formalization could be profound. Now, anyone or any bot can access the formal Stacks library and build new formal mathematics upon it. And together, the CFSG and Stacks formalizations represent a significant step toward formalizing all of mathematics. “They show the two ways formalization has to scale: one very long proof, and one very wide body of connected theory,” says Ono.

Ono and Voss see numerous advantages to a future in which formal mathematics becomes central: rubber stamping theorems and proofs, preserving knowledge as expertise dwindles over time, and building a shared library as a basis for creating new mathematics. They are also optimistic about humanity’s role when all of mathematics is formalized, viewing formalization as a launchpad for humans to use and understand mathematics in new and innovative ways. “Aren’t we excited to think more broadly about what the future of math can be?,” asks Ono. “I’m excited by that. I’m really excited.”

Related Articles Around the Web

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 IEEE Spectrum AI(Startup Makes Headway Towards Formalizing All of Math With AI),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗